Fetal heart rate signal feature extraction algorithm based on EMAU-Net network

Abstract

The invention discloses a fetal heart rate signal feature extraction algorithm based on an EMAU-Net network, and solves the technical problem of huge difference between a current computer automatic fetal heart rate signal analysis calculation result and expert judgment. According to the fetal heart rate signal feature extraction algorithm based on the EMAU-Net network, firstly, data preprocessing operation such as denoising and data enhancement is carried out on an input original fetal heart rate (FHR) signal, then the EMAU-Net network is used for segmenting an acceleration / deceleration area of the FHR signal, the segmented acceleration / deceleration area is eliminated, and the FHR signal feature extraction algorithm based on the EMAU-Net network is obtained. And calculating the average value of the remaining signals as the reference baseline of the FHR signal, and finally optimizing the judgment of the baseline, acceleration and deceleration to obtain the final baseline, acceleration and deceleration of the FHR signal. Compared with an existing method, the fetal heart rate signal feature extraction algorithm based on the EMAU-Net network has higher performance and has more advantages in the aspect of processing complex conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedical signal processing, and in particular to a fetal heart rate signal feature extraction algorithm based on an EMAU-Net network. Background Art

[0002] The main reason for the difference between the automatic feature analysis of fetal cardiotocography (CTG) and the expert consensus is the inconsistency between the results of the fetal heart rate baseline calculation and the interpretation between experts. The fetal heart rate baseline is the primary indicator of the CTG automatic feature analysis and is directly related to the accurate interpretation of acceleration and deceleration. Therefore, if the result of the baseline calculation is wrong, it will often cause a series of subsequent parameter analysis errors. At present, the mainstream baseline automatic calculation strategy uses filtering methods or progressive trimming methods to achieve the positioning of acceleration and deceleration to derive the baseline.

[0003] Filtering can eliminate the signal variability around the baseline. In addition, since acceleration and deceleration change faster than the signal baseline, filtering can remove most of the acceleration and deceleration, so the baseline can be obtained by filtering. For example: low-pass filtering (e.g., Cazares, Jimenez, Mantel, Pardey, Taylor), median filtering (e.g., Boudet), and other sliding filters (e.g., Campos, Madea, Mongelli). However, there are some problems with filtering methods: due to sudden acceleration / deceleration and low FHR values, low-pass filtering has boundary oscillation and acceleration / deceleration tailing problems, making it difficult to completely filter out acceleration / deceleration. When using median filtering and other sliding filters, it is easy to ignore insignificant acceleration / deceleration (e.g., long deceleration, stable acceleration / deceleration period longer than the baseline period, etc.). For complex situations, when acceleration / deceleration and baseline alternate, filtering determines that the baseline is the middle value of the FHR signal or the maximum frequency value within a limited window size, which may mislead subsequent baseline analysis.

[0004] The progressive trimming method can detect as many accelerations / decelerations as possible through multiple trimming, thereby calculating the baseline more accurately. However, this method is limited by the baseline value estimated for the first time. In the case of less acceleration / deceleration and shorter time, if the first calculated baseline is close to the "true" baseline, this method will help to obtain a more accurate baseline. However, in the case of more acceleration / deceleration and longer duration, if the first baseline estimation misjudges the baseline and acceleration / deceleration, it will not only fail to improve the baseline, but will aggravate the deviation. In addition, the baseline estimation step of progressive trimming mainly uses a filtering-based method, which means that the performance of the filter determines the necessity of progressive trimming. Therefore, if the accuracy of the acceleration / deceleration detected by the filter is high enough, the "true" baseline can be calculated without progressive trimming.

[0005] As mentioned above, these methods are highly dependent on limited filtering performance and are therefore not applicable to clinical FHR signals with random characteristics. Unlike automatic calculation methods, clinicians also consider the morphology and long-term correlation of the current signal when locating acceleration / deceleration, which is one of the reasons why there is a huge difference between the current computer automatic analysis results and expert judgment. Summary of the invention

[0006] The purpose of the present invention is to solve the above-mentioned defects in the prior art and provide a fetal heart rate signal feature extraction algorithm based on the EMAU-Net network.

[0007] The purpose of the present invention can be achieved by adopting the following technical solutions:

[0008] A fetal heart rate signal feature extraction algorithm based on EMAU-Net network, characterized in that the fetal heart rate signal feature extraction algorithm comprises the following steps:

[0009] S1, FHR signal preprocessing, including denoising and data enhancement;

[0010] S2, taking the FHR signal sub-segments obtained by the processing in step S1 as input, and using the EMAU-Net network to predict the acceleration / deceleration of these sub-segments, splicing all the predicted sub-segments to obtain the predicted label of the entire FHR signal;

[0011] S3. Based on the baseline calculation strategy of the long-term and short-term filters, the FHR signal obtained by the processing in step S2 is processed to finally obtain the baseline, acceleration and deceleration of the FHR signal.

[0012] Furthermore, the EMAU-Net is composed of four U-Nets (MAU-Nets) with multi-attention mechanisms, whose convolution kernel sizes (KS) are 21, 31, 61 and 81 respectively. By integrating the prediction results of the four MAU-Nets using an “or” operation, potential acceleration / deceleration point areas of the signal sub-segment can be detected;

[0013] The MAU-Net is based on the U-Net structure, wherein the U-Net structure consists of an encoder, a decoder, and a jump connection. The encoder includes two repeatedly used 1×KS convolutions, each followed by a batch normalization layer (BN), a rectified linear unit (Relu), and a 1×2 maximum pooling operation (Maxpooling 2), and downsampling with a stride of 2; each step in the decoder includes upsampling of the feature map, followed by a 1×2 convolution ("up convolution"), which reduces the number of feature channels by half, and jumps to the corresponding feature map of the encoder, as well as two 1×KS convolutions, each followed by a BN and a ReLu; in the last layer, a 1×1 convolution is used to map the 8-channel feature vector to the required 3 classes; in addition to the U-Net structure, an efficient channel attention module (ECA) and a non-local attention module (NLA) are added to the encoder of the MAU-Net structure;

[0014] The efficient channel attention module (ECA) inputs an H×C signal feature map, where C and H are the channel dimension and signal length respectively, and uses global average pooling (GAP) to obtain aggregated features. The channel weights are generated by performing fast one-dimensional convolution with a convolution kernel size of k and a Sigmoid function α. ​​After obtaining the weight of each feature channel, the weight is applied to each original feature channel, and the importance of different channels can be learned. The convolution kernel size k represents the coverage of local cross-channel interactions, that is, how many adjacent channels participate in the attention prediction of a channel. k is adaptively determined by the channel dimension C, and the formula is as follows:

[0015]

[0016] Where |t| odd represents the nearest odd number of t; γ and b are set to 2 and 1 respectively;

[0017] The non-local attention module (NLA) can directly capture long-range dependencies by calculating the interaction between any two positions, rather than being limited to adjacent points. It is equivalent to constructing a convolution kernel as large as the feature map size, so as to obtain more information. Let X be the input signal feature and Z be the output signal feature of the same size as X. The non-local operation can be defined as:

[0018]

[0019] Where i is the position index of Z, j is all possible position indexes obtained by traversing X, i,j∈H,W g ·X j It is to calculate the representation of X at position j, the function f(X i ,Xj ) is the calculation of X i Position and X j The similarity of positions uses the EmbeddedGaussian function to represent this pairwise relationship. C(X) is a normalization function that ensures that the overall information remains unchanged before and after the transformation. It is defined as:

[0020]

[0021] W u ·X i , W v ·X j The input is linearly mapped and a softmax calculation is performed to obtain pairwise relationships from it. By using these relationships and the representations of all positions on X, the value of position i is calculated, which is consistent with the input X i Add together to get the final output Z i .

[0022] Furthermore, a hybrid loss function is used, including the Cross-Entropy (CE) loss function, the Generalized-Dice (GD) loss function, and the Kullback–Leibler (KL) divergence, and its formula is as follows:

[0023] LOSS=0.99GD+0.01CE+0.01KL

[0024] The CE loss is used to evaluate the difference between the probability distribution obtained by the current training and the true distribution, while the KL is used to compare the proximity of the two probability distributions. The formula is as follows:

[0025]

[0026]

[0027] Where H is the length of the FHR signal, y n is the predicted value, a n is the true value, p represents the true distribution, and q represents the fitted distribution of p;

[0028] The GD loss integrates the Dice of multiple categories and uses an indicator to quantify the segmentation results. The formula is as follows:

[0029]

[0030] where r ln is the true value of category l at the nth time, and p ln is the corresponding predicted probability value, w l The weight for each category is as follows:

[0031]

[0032] Furthermore, the baseline calculation strategy based on the long-term and short-term filter is as follows:

[0033] First, based on the prediction, the preprocessed FHR signal x[n] and its predicted label y[n] are obtained, n = 1, 2, ..., N; then, the discrete baseline state signal L[n] is calculated as follows:

[0034]

[0035] Next, a long-term median filter is used to calculate the reference baseline L ref [n], as follows:

[0036]

[0037]

[0038] Where k is the window length of the filter, which is set to 20 minutes, s[n] is the percentage of baseline points in the sliding window, and points with s[n]≤0.1 are defined as invalid points and replaced by linear interpolation; next, with L ref In contrast, baseline points in L[n] that fluctuate more than 15 bpm are considered abnormal and removed. The calculation method is as follows:

[0039]

[0040] Then, the outliers in L'[n] that are 0 are processed by linear interpolation; finally, L'[n] passes through a short-term median filter with a window function length of 2 minutes to obtain the final baseline.

[0041] Compared with the prior art, the present invention has the following advantages and effects:

[0042] In baseline estimation, the problem of locating acceleration / deceleration is very similar to the segmentation problem in deep learning. On the one hand, the advantage of deep learning in capturing time series morphological information is used to improve the accuracy of locating acceleration and deceleration areas; on the other hand, based on the acceleration / deceleration positioning results, a baseline calculation strategy based on long-term and short-term filters is proposed, which combines the long-term and short-term prior knowledge of FHR signals with morphological characteristics to improve the accuracy of baseline calculation and get closer to the baseline interpretation method of clinicians. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of a fetal heart rate signal feature extraction algorithm based on the EMAU-Net network disclosed in an embodiment of the present invention;

[0044] Figure 2 is a schematic diagram of a calculation process of a fetal heart rate signal baseline disclosed in an embodiment of the present invention;

[0045] Figure 3 It is a schematic diagram of baseline, acceleration and deceleration (black rectangles indicate areas of false detection) of a 20-minute fetal heart rate signal of 90 private FHR signals from the Lille database in France disclosed in an embodiment of the present invention, which are annotated by an expert and the method of the present invention;

[0046] Figure 4 It is a schematic diagram of the baseline, acceleration and deceleration (black rectangles represent the areas that are falsely detected, and gray rectangles represent the areas that are not detected) of a 20-minute fetal heart rate signal annotated by an expert and the method of the present invention of 331 FHR signals from the Huayi feature analysis database disclosed in an embodiment of the present invention;

[0047] Figure 5 It is a schematic diagram of the performance of the method of the present invention disclosed in the embodiment of the present invention and the optimal method WMFB (Boudet et al., 2019) in evaluating complex cases. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] Example

[0050] This embodiment discloses a fetal heart rate signal feature extraction algorithm based on the EMAU-Net network, which specifically includes the following steps:

[0051] S1, FHR signal preprocessing, including denoising and data enhancement;

[0052] S2, taking the FHR signal sub-segments obtained by the processing in step S1 as input, and using the EMAU-Net network to predict the acceleration / deceleration of these sub-segments, splicing all the predicted sub-segments to obtain the predicted label of the entire FHR signal;

[0053] S3. Based on the baseline calculation strategy of the long-term and short-term filters, the FHR signal obtained by the processing in step S2 is processed to finally obtain the baseline, acceleration and deceleration of the FHR signal.

[0054] The model was trained and tested on a computer equipped with a 2.60GHz CPU, an NVIDIA GeForceRTX2080Ti GPU and 128GB of memory. The program development language was Python, using the Keras backend based on the TensorFlow framework.

[0055] The training set is 50 public FHR signals in the Lille database in France, and 2139 signal segments are obtained through data enhancement. The validation set is 16 public FHR signals in the Lille database in France, and 542 signal segments are obtained through data enhancement. The initial learning rate is set to 0.001, and the learning rate decays by 0.1 times every 50 cycles, for a total of 150 cycles. 90 private FHR signals of the Lille database in France and 331 FHR signals of the Chinese Medical Feature Analysis Database were used as two independent test sets respectively. Experiments were conducted on the 90 private FHR signals of the Lille database in France. Compared with the best method WMFB (Boudet et al., 2019), although the deceleration F-measure (Dec.F-measure) calculated by the method of the present invention was reduced by 1%, the acceleration F-measure (Acc.F-measure) was improved by 10%, the baseline root mean square deviation (BL.RMSD) was improved by 0.02bpm, the comprehensive inconsistency coefficient (SI) was improved by 0.4%, and the morphological analysis inconsistency index (MADI) was improved by 0.06%. Experiments were conducted on 331 FHR signals of the Chinese Medical Feature Analysis Database. Compared with the best method WMFB (Boudet et al., 2019), the best results were achieved in all indicators, verifying that the method of the present invention has higher performance than the existing methods.

[0056] Depend on Figure 5 It can be seen that when acceleration and baseline alternate, WMFB is more easily affected by acceleration than the method of the present invention, resulting in confusion between baseline and acceleration judgments; when a stable acceleration period is longer than the baseline period, WMFB is likely to judge the extended acceleration as the baseline, but the method of the present invention will not; when extended deceleration occurs, WMFB is prone to abnormal jumps in judging the baseline, while the method of the present invention is more stable. Therefore, compared with WMFB, the method of the present invention has more advantages in handling complex situations.

[0057] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A fetal heart rate signal feature extraction algorithm based on EMAU-Net network, characterized in that: The fetal heart rate signal feature extraction algorithm comprises the following steps: S1, FHR signal preprocessing, including denoising and data enhancement; S2, taking the FHR signal sub-segments obtained by the processing in step S1 as input, and using the EMAU-Net network to predict the acceleration / deceleration of these sub-segments, splicing all the predicted sub-segments to obtain the predicted label of the entire FHR signal; S3. Based on the baseline calculation strategy of the long-term and short-term filters, the FHR signal obtained by the processing in step S2 is processed to finally obtain the baseline, acceleration and deceleration of the FHR signal.

2. The fetal heart rate signal feature extraction algorithm based on the EMAU-Net network according to claim 1 is characterized in that: The EMAU-Net is composed of four U-Nets (MAU-Nets) with multi-attention mechanisms, whose convolution kernel sizes (KS) are 21, 31, 61 and 81 respectively. By integrating the prediction results of the four MAU-Nets using an "or" operation, the potential acceleration / deceleration point area of ​​the signal sub-segment can be detected; The MAU-Net is based on the U-Net structure, wherein the U-Net structure consists of an encoder, a decoder, and a jump connection. The encoder includes two repeated 1×KS convolutions, each followed by a batch normalization layer (BN), a rectified linear unit (Relu), and a 1×2 maximum pooling operation (Maxpooling 2), and downsampling with a stride of 2; each step in the decoder includes upsampling of the feature map, followed by a 1×2 convolution ("up convolution"), which reduces the number of feature channels by half, and jumps to the corresponding feature map of the encoder, as well as two 1×KS convolutions, each followed by a BN and a ReLu; in the last layer, a 1×1 convolution is used to map the 8-channel feature vector to the required 3 classes; in addition to the U-Net structure, an efficient channel attention module (ECA) and a non-local attention module (NLA) are added to the encoder of the MAU-Net structure; The efficient channel attention module (ECA) inputs an H×C signal feature map, where C and H are the channel dimension and signal length respectively, and uses global average pooling (GAP) to obtain aggregated features. The channel weights are generated by performing fast one-dimensional convolution with a convolution kernel size of k and a Sigmoid function α. ​​After obtaining the weight of each feature channel, the weight is applied to each original feature channel, and the importance of different channels can be learned. The convolution kernel size k represents the coverage of local cross-channel interactions, that is, how many adjacent channels participate in the attention prediction of a channel. k is adaptively determined by the channel dimension C, and the formula is as follows: Where |t| odd represents the nearest odd number of t; γ and b are set to 2 and 1 respectively; The non-local attention module (NLA) can directly capture long-range dependencies by calculating the interaction between any two positions, rather than being limited to adjacent points. It is equivalent to constructing a convolution kernel as large as the feature map size, so as to obtain more information. Let X be the input signal feature and Z be the output signal feature of the same size as X. The non-local operation can be defined as: Where i is the position index of Z, j is all possible position indexes obtained by traversing X, i,j∈H,W g ·X j It is to calculate the representation of X at position j, the function f(X i ,X j ) is the calculation of X i Position and X j The similarity of positions uses the Embedded Gaussian function to represent this pairwise relationship. C(X) is a normalization function that ensures that the overall information remains unchanged before and after the transformation. It is defined as: W u ·X i , W v ·X j The input is linearly mapped and a softmax calculation is performed to obtain pairwise relationships from it. By using these relationships and the representations of all positions on X, the value of position i is calculated, which is consistent with the input X i Add together to get the final output Z i .

3. The fetal heart rate signal feature extraction algorithm based on the EMAU-Net network according to claim 1, characterized in that: A hybrid loss function is used, including the Cross-Entropy (CE) loss function, the Generalized-Dice (GD) loss function, and the Kullback–Leibler (KL) divergence, and its formula is as follows: LOSS=0.99GD+0.01CE+0.01KL The CE loss is used to evaluate the difference between the probability distribution obtained by the current training and the true distribution, while the KL is used to compare the proximity of the two probability distributions. The formula is as follows: Where H is the length of the FHR signal, y n is the predicted value, a n is the true value, p represents the true distribution, and q represents the fitted distribution of p; The GD loss integrates the Dice of multiple categories and uses an indicator to quantify the segmentation results. The formula is as follows: where r ln is the true value of category l at the nth time, and p ln is the corresponding predicted probability value, w l The weight for each category is as follows:

4. The fetal heart rate signal feature extraction algorithm based on the EMAU-Net network according to claim 1, characterized in that: The baseline calculation strategy based on the long-term and short-term filter is as follows: First, based on the prediction, the preprocessed FHR signal x[n] and its predicted label y[n] are obtained, n = 1, 2, ..., N; then, the discrete baseline state signal L[n] is calculated as follows: Next, a long-term median filter is used to calculate the reference baseline L ref [n], as follows: Where k is the window function length of the filter, which is set to 20 minutes, s[n] is the percentage of baseline points in the sliding window, and points with s[n]≤0.1 are defined as invalid points and replaced by linear interpolation; Next, with L ref In contrast, baseline points in L[n] that fluctuate more than 15 bpm are considered abnormal and removed. The calculation method is as follows: Then, the outliers in L'[n] that are 0 are processed by linear interpolation; finally, L'[n] passes through a short-term median filter with a window function length of 2 minutes to obtain the final baseline.

Images